engine/core: unify the three trick-winner computation call paths

[valmathieu]

Problem

The trick-winner rule is computed through three independent call paths that all delegate to the same core current_winner free function (packages/contrai-core/src/contrai_core/trick.py):

1. PlayState.trick_winners maps current_winner over completed_tricks (packages/contrai-core/src/contrai_core/play.py).
2. Round recomputes the winner on its mutable mirror Trick (packages/contrai-engine/src/contrai_engine/model/round/round.py, get_current_winner on current_trick) — this is what feeds team_tricks and last_trick_winner for scoring.
3. The view builds a throwaway proxy Trick and pushes directly onto proxy.plays to reuse the rule (packages/contrai-engine/src/contrai_engine/view/state_helpers.py, _current_winner).

They agree today because they share the underlying function, but three call sites independently asserting the same fact is redundancy waiting to drift — especially the view's proxy-object construction, which bypasses Trick's normal API.

Proposal

Unify on the core Play records / PlayState as the single source of truth:

• Round reads winners from PlayState.trick_winners instead of recomputing on its mirror Trick.
• The view's state_helpers._current_winner consumes the same plays-based rule directly (no proxy Trick).
• Longer term this folds into reducing the Round mirror-state duplication (Trick holding (Player, Card) tuples vs core Play records).

Notes

• Pure refactor: no behaviour change; existing round-lifecycle and view tests must stay green.
• Good to schedule alongside or after the Position migration since both touch Round/view seams, but not blocked by it.